基于大数据方法的玄武岩构造环境智能确定与数据挖掘

Basalt discrimination diagrams have been widely used for determining tectonic settings. Since the first basalt discrimination diagram was proposed by Pearce in 1971, dozens of discrimination diagrams have emerged. However, the information in a discrimination diagram is usually 2 ~ 3 elements, and the amount of samples for designing a discrimination diagram was usually small, leading to a limitation of their applications. To improve the effectiveness and accuracy of determination, in this study, a set of methods based on intelligent algorithms and chemical composition of basalts is presented. The samples used in this research comprise 3 kinds of basalts: mid-ocean ridge basalts (MORB), ocean island basalts (OIB) and island arc basalts (IAB). The amount of the samples analyzed is 755. At first, three trace elements discrimination diagrams and two major elements discrimination diagrams, including TiO₂-MnO-P₂O₅ diagram, FeO^T-MgO-Al₂O₃diagram, Ti-Zr-Y diagram, Zr/Y-Zr diagram and Ti-Zr diagram, are adopted for plotting the samples. Considering the limitations of the diagrams, the samples should be filtered before being plotted. The results show that the Zr/Y-Zr diagram can reach a high accuracy of 90% with the filtered samples. However, its accuracy is less than 75% when using the whole samples. In this paper, the methods of Naive Bayes (NB), K-Nearest Neighbors (KNN), Support Vector Machine (SVM) and Random Forest (RF) are adopted for determination. In training, every sample is represented by a 51-dimension vector that comprises 11 major elements, 35 trace elements and 5 isotopes, and they are not filtered. It shows that the worst has more than 75% of accuracy. The best result is made by RF, and its training accuracy is 100%. In the advanced analysis, the results show that the RF can reach a high validation accuracy of 88.46%. To improve the practicability of intelligent algorithms, the Bayes theorem is used to calculate the inverse probabilities. After that, by simulating data missing, the robust of the algorithms are verified, and it shows that RF and NB are the best. Finally, by extracting the decision trees of RF algorithm, the importance of the 51 features of samples are calculated, and then the major elements and trace elements that affect the determination most are found out. In conclusion, it is more effective, accurate and functional to determine tectonic settings by intelligent algorithms, and this set of method is worthy of promotion.